Abstract
Pretrained deep learning models offer the potential to accelerate the analysis of prostate multiparametric MRI across clinical settings. Yet, before such models can be translated into broader use, their robustness to substantial differences in patient populations, scanner hardware, and imaging protocols must be understood. We externally evaluate three publicly available pretrained prostate MRI models on 881 patients from a tertiary imaging centre in Ghana, acquired using a 1.5T scanner. In the absence of comprehensive expert annotations, we propose a proxy validation framework based on geometric plausibility, inter model agreement, prediction uncertainty, and threshold sensitivity. We also compare acquisition characteristics with major public datasets to identify potential sources of domain shift. The models showed markedly different deployment behaviour. MedSAM produced constrained predictions with limited spatial selectivity, DIAG nnU-Net generated sparse and anatomically inconsistent detections, while ProFound produced widespread high confidence predictions with frequent anatomical overextension. Low inter model agreement persisted across probability thresholds. The Ghanaian cohort also differed in diffusion imaging characteristics, including a higher b value of (1400 s/mm2) and a shifted ADC intensity distribution. The results suggest that pretrained prostate MRI models may behave substantially differently when applied to geographically and technically distinct imaging data. Our proxy validation framework provides a practical approach for characterising model behaviour in settings where comprehensive expert annotations are unavailable.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_029.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=WFBNjADIuI
BibTex
@InProceedings{AdjPri_Domain_MICCAISAT2026,
author = { Adjei, Prince Ebenezer AND Frimpong, George Asafu Adjaye AND Amuasi, John},
title = { { Domain Shift in Prostate MRI AI: Proxy Validation on a Low-Resource 1.5T Clinical Cohort } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}